<img width="2318" height="1448" alt="image" src="https://github.com/user-attachments/assets/bca71d69-4972-420c-b14e-c8cf74a17de9" /> ## Summary Everything a knowledge base could not read, and the pieces it cut them into saying what they are. **Reading a PDF.** A PDF records glyphs at coordinates; whether they are a heading, a table cell or a running footer is inferred from how they look. There are now two readers. The geometric one infers structure from type size and position and needs nothing but the file. The other asks deepdoc's models, ported into `raven/knowledge/parser/deepdoc`: a layout model that says what each region is, a detector and recognizer that read a page carrying no text at all, and a table-structure model that rebuilds a grid from a picture of one. Detecting boxes around text the file already spells out exactly costs an inference pass per page to learn less, and recognising it costs accuracy. The layout graph answers with a shape its own postprocessing raises on, and its classes are DocLayout-YOLO's order rather than the list Both were established by running the model, and the label order is pinned by a test. **Tables go in as tables.** `<table>` is the only shape that keeps what a grid states -- which column a cell stood under, which cells are merged, which row is the header -- and it is the shape a model reads a table in, which is what these chunks are for. PDF, Word and PowerPoint all state their merges exactly, and they now survive as `colspan` and `rowspan`. The flat form this replaces had to repeat a header standing over two columns across both of them, because there was nowhere to record a span; it had no callers left afterwards and is gone. Two deviations from upstream, both about what the string is for: every cell's text is escaped, where upstream interpolates it raw and a cell holding a `<` produces markup that is no longer a table, and the span attributes are written only where there is a span, where upstream emits `<td >` for every ordinary cell. **A picture of where each chunk came from.** A chunk is a few hundred words with a page number on it, which is enough to cite and not enough to answer whether the parser read the page correctly. So every PDF chunk that knows where it sits carries one image: the regions it covers, cropped out of the rendered pages and stacked. here they are already structured in the parser's element spans, so the markers would only put text nobody wrote into the thing being embedded. The panel draws a chunk's tables as tables, parsed and rebuilt as elements rather than injected: a person can rewrite any piece from that same panel, so what comes back from the store is user input however it got there. **The page that reads them.** Upstream retired the knowledge page while this branch was out; it is rebuilt against the new architecture, and the half that never existed is the one a reader needs -- what a document actually became once it was cut and embedded. Documents sit in one-level folders and arrive by four routes: a file, a folder, a URL, a written note. Adding and indexing are two calls on the contract and nothing drains that queue on its own, so a page that stopped after the first left every upload saying Queued for good. Opening a file takes the whole page: the document on the left, the pieces the index holds of it on the right. The left half is the file drawn as its own pages rather than framed by the browser's PDF viewer, which reads where to open once and answers to nothing afterwards -- and which nothing can be drawn on. Pages are rendered once and kept, and the region a piece was cut from is washed over the page it came from, placed in fractions of the page so the picture's own scale never enters the arithmetic. Clicking a piece takes the file there; double-clicking opens it for rewriting. The pieces can be searched, filtered by state, turned on and off, deleted, rewritten and added to, one at a time or in a batch. Every one of those was already on the contract with no reader for it. A base's settings are back and fuller: how many pieces a search returns, how documents are cut, and how much prose comes with a table or a figure. The embedding model is a picker while the base holds no pieces and a stated fact once it does -- moving it then is the collection made again at the new width and every document re-cut, which is not a thing a panel does by being saved. **An agent can read them.** The retrieval engine was there and nothing could reach it: no agent had a tool that opened a library. A reader now points a conversation at bases from the composer, the way they pin a working directory, and the selection rides in the same place -- the session's metadata, bound around the turn. The tool takes a query and never a base. The scope is the reader's choice, so a model cannot reach material nobody offered it, and it is withheld where nothing was picked rather than offered and refused. A hit says which document and page it came from, because a hit a reader cannot place is a hit they cannot check -- and the passages are pressable in the transcript, walking into the base, the document, the folder it is filed under and then the piece. A citation a reader cannot follow is a claim they have to take on trust, and the whole point of searching their own material is that they can go and look. **The word segmenter and the resources.** `raven.core.Tokenizer` is Infinity's `rag_tokenizer`, ported: a forward and backward maximum match over a dictionary trie, with stemming and lemmatisation on the latin path. XGBoost is not carried over. None of the data lives in git -- the dictionary, the NLTK corpora and the ONNX graphs run to about a hundred megabytes, so `make fetch-resources` on an install and an optional Docker layer behind `RAVEN_RESOURCES` both point at one script. A checkout that could not reach the network still works, and every feature that reads this data says so when asked. ## Type - [ ] Fix - [x] Feature - [ ] Docs - [ ] CI / tooling - [ ] Refactor - [ ] Other ## Verification Run on macOS against Python 3.12, with the deepdoc weights fetched (the model tests skip without them): ``` uv run pytest 27450 passed, 111 skipped, 3 failed npm run test --prefix ui-web 3276 passed npm run test --prefix ui-tui 2106 passed npm run type-check --prefix ui-web clean (tsc) npm run lint --prefix ui-web clean (4 pre-existing hook warnings) make lint-python clean (ruff check + format) make lint-types clean (ty) make lint-imports 10 contracts kept, 0 broken make check-source-language clean make check-large-files clean make build-ui built, both boot snapshots match node ui-web/scripts/check-css.mjs clean node ui-web/scripts/check-class-namespace.mjs clean ``` The three Python failures are not this branch's and were each confirmed on the base alone: one asks a host without LibreOffice to refuse a render and this host has LibreOffice installed, and two simulation cases fail on an unmodified checkout of the base commit. The RPC contract was regenerated for both consumers and both match it (`npm run gen --prefix ui-web`, `npm run gen:rpc --prefix ui-tui`, then `gen:check`). The page's boot snapshots and region goldens were refreshed where the composer bar gained a button, and each records that button and nothing else. - [x] Relevant tests pass locally - [x] Relevant lint / type checks pass locally - [x] User-facing docs or screenshots are updated when needed ## Risk A PDF's tables, and a Word or PowerPoint file's, are now indexed as HTML rather than as header-prefixed lines. Chunk text changes shape for those documents, so a base indexed before this and searched after it holds both forms until its documents are reindexed; retrieval keeps working either way, because both forms are text. The same applies to how much prose a table or a figure carries with it: the setting now reaches as far as it says, where before it stopped at the nearest paragraph, so a chunk cut before this keeps the narrower context until its document is rebuilt. Rolling back is reverting the branch and reindexing. The deepdoc models are optional. An install that never ran `make fetch-resources` reads PDFs geometrically, exactly as before, and says so when the models are asked for by name. Nothing here reaches the network at request time. Chunk crops and rendered pages are written at index time and on first view, both under the state directory rather than in the repository, and both are deleted with their document. A page is immutable for the life of its document id, so the route serving them caches; the document route does not. All three routes take ids and no path, and the two image routes answer under the strictest sandbox in the tree. What an agent may search is the reader's choice and nothing else. The tool takes no base argument, so a model cannot name one, and it is absent from the schema entirely for a conversation that named none -- a lane nobody pointed at a base is unchanged by this branch. The bases a session carries are stored beside its working directory and travel no further. - [x] Security impact considered - [x] Backward compatibility considered - [x] Rollback path is clear for risky changes ## Related Issues N/A --------- Co-authored-by: Claude (claude-opus-5) <noreply@anthropic.com>
445 lines
16 KiB
Python
445 lines
16 KiB
Python
"""The PDF parser: the text layer where there is one, a model's eyes where not."""
|
|
|
|
from __future__ import annotations
|
|
|
|
import asyncio
|
|
|
|
import pytest
|
|
|
|
from raven.knowledge._naive_chunker import NaiveChunker
|
|
from raven.knowledge._notes import collecting
|
|
from raven.knowledge.parser import ELEMENTS, LAYOUT_TYPE, PAGE_NUMBER, LayoutType
|
|
from raven.knowledge.parser import pdf_parser as pdf_module
|
|
from raven.knowledge.parser.pdf_parser import PdfParser
|
|
|
|
|
|
def _pdf(build) -> bytes:
|
|
"""A PDF built page by page, as bytes."""
|
|
import pymupdf
|
|
|
|
document = pymupdf.open()
|
|
build(document, pymupdf)
|
|
raw = document.tobytes()
|
|
document.close()
|
|
return raw
|
|
|
|
|
|
def _prose(document, pymupdf, *, heading: str = "", body: "list[str] | None" = None) -> None:
|
|
page = document.new_page()
|
|
at = 80
|
|
if heading:
|
|
page.insert_text(pymupdf.Point(60, at), heading, fontsize=20)
|
|
at += 40
|
|
for paragraph in body or []:
|
|
page.insert_text(pymupdf.Point(60, at), paragraph, fontsize=10)
|
|
at += 24
|
|
|
|
|
|
def _ruled_table(page, pymupdf, rows: "list[list[str]]", *, top: float = 300.0) -> None:
|
|
"""A table `find_tables` can see: it keys on the ruling lines."""
|
|
xs = [60 + 140 * n for n in range(len(rows[0]) + 1)]
|
|
ys = [top + 30 * n for n in range(len(rows) + 1)]
|
|
for y in ys:
|
|
page.draw_line(pymupdf.Point(xs[0], y), pymupdf.Point(xs[-1], y))
|
|
for x in xs:
|
|
page.draw_line(pymupdf.Point(x, ys[0]), pymupdf.Point(x, ys[-1]))
|
|
for r, row in enumerate(rows):
|
|
for c, value in enumerate(row):
|
|
page.insert_text(pymupdf.Point(xs[c] + 6, ys[r] + 20), value, fontsize=11)
|
|
|
|
|
|
class _Vision:
|
|
"""A page describer that answers without a network."""
|
|
|
|
max_figures = 64
|
|
|
|
def __init__(self, text: str = "A scanned invoice for 1,200 euro.", error: Exception | None = None) -> None:
|
|
self.text = text
|
|
self.error = error
|
|
self.seen: list[int] = []
|
|
|
|
async def describe(self, image: bytes, *, mime: str = "", context_above: str = "", context_below: str = "") -> str:
|
|
self.seen.append(len(image))
|
|
if self.error is not None:
|
|
raise self.error
|
|
return self.text
|
|
|
|
|
|
def _parse(raw: bytes, model=None, name: str = "report.pdf"):
|
|
return asyncio.run(PdfParser(model).parse(raw, name))
|
|
|
|
|
|
# -- what it claims ------------------------------------------------
|
|
|
|
|
|
def test_it_claims_pdf_and_nothing_else() -> None:
|
|
assert PdfParser.supported_media_types == ["application/pdf"]
|
|
assert PdfParser.supported_extensions() == [".pdf"]
|
|
|
|
|
|
# -- the text layer ------------------------------------------------
|
|
|
|
|
|
def test_a_heading_bounds_a_section() -> None:
|
|
"""The same cut the Word and Markdown parsers make, so a document exported
|
|
from one format and indexed from the other lands the same way."""
|
|
raw = _pdf(lambda d, m: _prose(d, m, heading="Quarterly report", body=["Revenue held up.", "Costs did not."]))
|
|
|
|
sections = _parse(raw)
|
|
|
|
assert len(sections) == 1
|
|
assert sections[0].metadata["heading_path"] == ["Quarterly report"]
|
|
assert "Revenue held up." in sections[0].content.text
|
|
|
|
|
|
def test_two_headings_are_two_sections() -> None:
|
|
def build(document, pymupdf):
|
|
_prose(document, pymupdf, heading="First part", body=["Alpha text here."])
|
|
_prose(document, pymupdf, heading="Second part", body=["Beta text here."])
|
|
|
|
sections = _parse(_pdf(build))
|
|
|
|
assert [s.metadata["heading_path"] for s in sections] == [["First part"], ["Second part"]]
|
|
|
|
|
|
def test_every_element_says_which_page_it_came_from() -> None:
|
|
"""A PDF is read to be cited, and a hit with no page is a hit a reader has
|
|
to go looking for."""
|
|
|
|
def build(document, pymupdf):
|
|
_prose(document, pymupdf, heading="Report", body=["Page one text."])
|
|
_prose(document, pymupdf, body=["Page two text."])
|
|
|
|
sections = _parse(_pdf(build))
|
|
|
|
pages = {span[PAGE_NUMBER] for section in sections for span in section.metadata[ELEMENTS]}
|
|
assert pages == {1, 2}
|
|
|
|
|
|
def test_an_element_carries_the_box_it_was_drawn_in() -> None:
|
|
raw = _pdf(lambda d, m: _prose(d, m, heading="Report", body=["Some prose."]))
|
|
|
|
span = _parse(raw)[0].metadata[ELEMENTS][0]
|
|
|
|
box = span["bbox"]
|
|
assert box["x0"] > 0 and box["x1"] > box["x0"]
|
|
assert box["bottom"] > box["top"], "measured downward from the top of the page, as a PDF states it"
|
|
|
|
|
|
def test_the_lines_of_a_paragraph_are_joined_into_one() -> None:
|
|
"""A PDF breaks a paragraph into a line per printed line. Kept as breaks,
|
|
one sentence indexes as several fragments."""
|
|
long_line = "This sentence is long enough that the page has to break it across two printed lines to fit."
|
|
raw = _pdf(lambda d, m: _prose(d, m, heading="Report", body=[long_line]))
|
|
|
|
text = _parse(raw)[0].content.text
|
|
|
|
assert long_line.split(".")[0] in text.replace("\n", " ")
|
|
|
|
|
|
# -- tables --------------------------------------------------------
|
|
|
|
|
|
def test_a_table_is_one_element_holding_the_whole_grid() -> None:
|
|
"""A PDF states where its cell boundaries are, and `<table>` is the shape
|
|
that keeps them. Flattened to lines a cell that spans two columns has
|
|
nowhere to go, and a header standing over a group of columns has nowhere
|
|
at all -- and a grid is what a model reads a table as."""
|
|
|
|
def build(document, pymupdf):
|
|
page = document.new_page()
|
|
page.insert_text(pymupdf.Point(60, 80), "Revenue", fontsize=20)
|
|
_ruled_table(page, pymupdf, [["Region", "Revenue"], ["EU", "1.2M"], ["US", "2.4M"]])
|
|
|
|
sections = _parse(_pdf(build))
|
|
|
|
tables = [span for s in sections for span in s.metadata[ELEMENTS] if span[LAYOUT_TYPE] == LayoutType.TABLE]
|
|
assert len(tables) == 1
|
|
body = sections[0].content.text
|
|
assert "<table>" in body and "</table>" in body
|
|
# The column names as a header row, and the same row whichever reader ran:
|
|
# the models are optional, and a document that indexed one way with them
|
|
# and another way without would make a base depend on the machine that
|
|
# built it.
|
|
assert "<tr><th>Region</th><th>Revenue</th></tr>" in body
|
|
assert "<tr><td>EU</td><td>1.2M</td></tr>" in body
|
|
|
|
|
|
def test_a_header_row_over_a_numeric_body_is_marked_as_one() -> None:
|
|
"""Which row is the header is a guess, and RAGFlow's guess is the one kept:
|
|
a row that breaks the body's dominant kind. So it fires where the body is
|
|
numbers and the header is words, and stays quiet on a grid of labels --
|
|
marking an arbitrary row as the header would be worse than marking none."""
|
|
|
|
def build(document, pymupdf):
|
|
page = document.new_page()
|
|
page.insert_text(pymupdf.Point(60, 80), "Results", fontsize=20)
|
|
_ruled_table(page, pymupdf, [["Region", "Q1"], ["101", "1.2"], ["202", "2.4"], ["303", "3.6"]])
|
|
|
|
body = _parse(_pdf(build))[0].content.text
|
|
|
|
assert "<tr><th>Region</th><th>Q1</th></tr>" in body
|
|
assert "<tr><td>101</td><td>1.2</td></tr>" in body
|
|
|
|
|
|
def test_a_table_cell_is_not_also_indexed_as_prose() -> None:
|
|
"""The cells are text blocks too. Indexed twice, a search for a value
|
|
answers with a line that has no column name attached."""
|
|
|
|
def build(document, pymupdf):
|
|
page = document.new_page()
|
|
page.insert_text(pymupdf.Point(60, 80), "Revenue", fontsize=20)
|
|
_ruled_table(page, pymupdf, [["Region", "Revenue"], ["EU", "1.2M"]])
|
|
|
|
text = _parse(_pdf(build))[0].content.text
|
|
|
|
assert text.count("1.2M") == 1
|
|
|
|
|
|
# -- a page with no text layer -------------------------------------
|
|
|
|
|
|
def _scanned(document, pymupdf) -> None:
|
|
"""A page with a picture on it and no text at all, as a scan is."""
|
|
page = document.new_page()
|
|
pixmap = pymupdf.Pixmap(pymupdf.csRGB, pymupdf.IRect(0, 0, 40, 40))
|
|
pixmap.set_rect(pixmap.irect, (240, 240, 240))
|
|
page.insert_image(pymupdf.Rect(60, 60, 400, 400), pixmap=pixmap)
|
|
|
|
|
|
def test_a_page_with_no_text_is_read_by_the_model() -> None:
|
|
"""The only path to what a scan says, and the reason there is no OCR
|
|
engine here."""
|
|
vision = _Vision()
|
|
|
|
sections = _parse(_pdf(_scanned), vision)
|
|
|
|
assert vision.seen, "the page was rendered and sent"
|
|
assert "scanned invoice" in sections[0].content.text
|
|
assert sections[0].metadata[ELEMENTS][0][PAGE_NUMBER] == 1
|
|
|
|
|
|
def test_only_the_pages_that_need_it_are_sent() -> None:
|
|
"""A model call per page of a readable document would be a bill for
|
|
nothing: the text layer is already there."""
|
|
|
|
def build(document, pymupdf):
|
|
_prose(document, pymupdf, heading="Report", body=["This page has text on it."])
|
|
_scanned(document, pymupdf)
|
|
|
|
vision = _Vision()
|
|
sections = _parse(_pdf(build), vision)
|
|
|
|
assert len(vision.seen) == 1, "the page with text was not sent"
|
|
assert "scanned invoice" in " ".join(s.content.text for s in sections)
|
|
|
|
|
|
def test_no_vision_model_leaves_a_note_rather_than_a_silent_gap(monkeypatch) -> None:
|
|
"""A scanned page indexed as nothing looks exactly like a page that said
|
|
nothing."""
|
|
monkeypatch.setattr("raven.knowledge._vision.load_vision_model", lambda: None)
|
|
|
|
with collecting() as notes:
|
|
sections = _parse(_pdf(_scanned))
|
|
|
|
assert sections == []
|
|
assert any("no vision model is configured" in n for n in notes)
|
|
|
|
|
|
def test_a_page_the_model_refuses_costs_only_that_page() -> None:
|
|
def build(document, pymupdf):
|
|
_prose(document, pymupdf, heading="Report", body=["This page has text on it."])
|
|
_scanned(document, pymupdf)
|
|
|
|
with collecting() as notes:
|
|
sections = _parse(_pdf(build), _Vision(error=RuntimeError("rate limited")))
|
|
|
|
assert "This page has text on it." in sections[0].content.text
|
|
assert any("rate limited" in n for n in notes)
|
|
|
|
|
|
# -- what it refuses -----------------------------------------------
|
|
|
|
|
|
def test_bytes_that_are_not_a_pdf_say_so() -> None:
|
|
with pytest.raises(ValueError, match="as a PDF"):
|
|
_parse(b"this is not a pdf at all")
|
|
|
|
|
|
def test_a_password_protected_file_says_which(tmp_path) -> None:
|
|
"""Not "could not parse": the reader can do something about a password."""
|
|
import pymupdf
|
|
|
|
document = pymupdf.open()
|
|
document.new_page().insert_text(pymupdf.Point(60, 80), "secret")
|
|
path = tmp_path / "locked.pdf"
|
|
document.save(str(path), encryption=pymupdf.PDF_ENCRYPT_AES_256, owner_pw="o", user_pw="u")
|
|
document.close()
|
|
|
|
with pytest.raises(ValueError, match="password-protected"):
|
|
asyncio.run(PdfParser().parse(str(path), "locked.pdf"))
|
|
|
|
|
|
def test_a_path_that_is_not_there_is_not_a_parse_failure(tmp_path) -> None:
|
|
with pytest.raises(FileNotFoundError):
|
|
asyncio.run(PdfParser().parse(str(tmp_path / "gone.pdf"), "gone.pdf"))
|
|
|
|
|
|
# -- end to end ----------------------------------------------------
|
|
|
|
|
|
def test_a_parsed_pdf_chunks_with_its_positions_intact() -> None:
|
|
"""The contract the whole thing is for: a chunk of this document can still
|
|
be pointed at the page it came from."""
|
|
|
|
def build(document, pymupdf):
|
|
_prose(document, pymupdf, heading="Handbook", body=["Alpha " * 40])
|
|
_prose(document, pymupdf, heading="Appendix", body=["Beta " * 40])
|
|
|
|
sections = _parse(_pdf(build))
|
|
chunks = asyncio.run(NaiveChunker(chunk_size=4096).chunk(sections))
|
|
|
|
assert len(chunks) == 1, "both sections fit one chunk, which is what merging metadata is for"
|
|
spans = chunks[0].metadata[ELEMENTS]
|
|
assert {span[PAGE_NUMBER] for span in spans} == {1, 2}
|
|
assert chunks[0].metadata["page_number"] == 1
|
|
assert chunks[0].metadata["page_end"] == 2
|
|
|
|
|
|
def test_a_sparse_page_is_not_mistaken_for_a_scan() -> None:
|
|
"""A section divider has nine characters on it and no picture. Sent to a
|
|
model, it costs a call to be told what the text layer already said."""
|
|
|
|
def build(document, pymupdf):
|
|
page = document.new_page()
|
|
page.insert_text(pymupdf.Point(60, 300), "Chapter 3", fontsize=24)
|
|
|
|
vision = _Vision()
|
|
sections = _parse(_pdf(build), vision)
|
|
|
|
assert vision.seen == [], "nothing was sent"
|
|
assert "Chapter 3" in sections[0].content.text
|
|
|
|
|
|
def test_a_picture_page_with_a_caption_is_still_read_as_a_scan() -> None:
|
|
"""The other half of the rule: a scan routinely carries a watermark or a
|
|
header from whatever produced it, and indexing those instead of the page
|
|
is worse than indexing nothing."""
|
|
|
|
def build(document, pymupdf):
|
|
_scanned(document, pymupdf)
|
|
document[0].insert_text(pymupdf.Point(60, 500), "page 1", fontsize=8)
|
|
|
|
vision = _Vision()
|
|
sections = _parse(_pdf(build), vision)
|
|
|
|
assert vision.seen, "the page was sent"
|
|
assert "scanned invoice" in sections[0].content.text
|
|
|
|
|
|
# -- what it does to the loop it is awaited on ---------------------
|
|
|
|
|
|
def test_reading_the_pages_happens_off_the_event_loop() -> None:
|
|
"""`parse` is awaited by `knowledge.documents.index` on the gateway's own
|
|
loop -- the one serving every socket, every turn and the heartbeat.
|
|
|
|
Reading a page is PyMuPDF rendering and ONNX inference with no await in
|
|
it. Run inline it stops all of that for as long as the document takes:
|
|
measured before this moved, a 40-page file blocked the loop for 21
|
|
seconds and answered no RPC in that time.
|
|
"""
|
|
import threading
|
|
|
|
seen: list[int] = []
|
|
real = pdf_module._page_elements
|
|
|
|
def recording(page, number, body):
|
|
seen.append(threading.get_ident())
|
|
return real(page, number, body)
|
|
|
|
raw = _pdf(lambda d, m: [_prose(d, m, heading=f"H{n}", body=["text"]) for n in range(3)])
|
|
|
|
async def main() -> int:
|
|
return threading.get_ident()
|
|
|
|
loop_thread = asyncio.run(main())
|
|
with pytest.MonkeyPatch.context() as patch:
|
|
patch.setattr(pdf_module, "_page_elements", recording)
|
|
asyncio.run(PdfParser(layout=False).parse(raw, "report.pdf"))
|
|
|
|
assert seen, "the pages were read"
|
|
assert all(ident != loop_thread for ident in seen), "read on a worker, not on the loop"
|
|
|
|
|
|
def test_the_loop_keeps_running_while_a_document_is_read() -> None:
|
|
"""The same property from the other side: something else awaiting on that
|
|
loop goes on being served.
|
|
|
|
The page reader is made to take a measurable moment, because the question
|
|
is whether the loop runs *during* a slow read -- a document fast enough to
|
|
finish between two ticks would pass either way.
|
|
"""
|
|
import threading
|
|
|
|
def slow_reader(page, number):
|
|
threading.Event().wait(0.05)
|
|
return None
|
|
|
|
raw = _pdf(lambda d, m: [_prose(d, m, heading=f"H{n}", body=["text"]) for n in range(4)])
|
|
|
|
async def main() -> int:
|
|
ticks = 0
|
|
|
|
async def ticker() -> None:
|
|
nonlocal ticks
|
|
while True:
|
|
await asyncio.sleep(0.005)
|
|
ticks += 1
|
|
|
|
with pytest.MonkeyPatch.context() as patch:
|
|
patch.setattr(PdfParser, "_layout_reader", lambda self: slow_reader)
|
|
beat = asyncio.create_task(ticker())
|
|
await asyncio.sleep(0)
|
|
await PdfParser().parse(raw, "report.pdf")
|
|
beat.cancel()
|
|
return ticks
|
|
|
|
# Four pages at 50ms is 200ms of reading; a loop that is free is ticking
|
|
# every 5ms through all of it, and a blocked one answers nothing.
|
|
assert asyncio.run(main()) > 10, "the loop was served while the document was read"
|
|
|
|
|
|
def test_two_documents_do_not_run_two_models_at_once() -> None:
|
|
"""Two uploads indexing together otherwise run two ONNX sessions over the
|
|
same cores, which makes both slower than running them in turn and leaves
|
|
nothing for the turns the gateway is serving beside them."""
|
|
import threading
|
|
|
|
inside = 0
|
|
overlapped = False
|
|
|
|
def slow_reader(page, number):
|
|
nonlocal inside, overlapped
|
|
inside += 1
|
|
overlapped = overlapped or inside > 1
|
|
try:
|
|
threading.Event().wait(0.01)
|
|
finally:
|
|
inside -= 1
|
|
return None
|
|
|
|
raw = _pdf(lambda d, m: [_prose(d, m, heading=f"H{n}", body=["text"]) for n in range(4)])
|
|
|
|
async def main() -> None:
|
|
with pytest.MonkeyPatch.context() as patch:
|
|
patch.setattr(PdfParser, "_layout_reader", lambda self: slow_reader)
|
|
await asyncio.gather(
|
|
PdfParser().parse(raw, "a.pdf"),
|
|
PdfParser().parse(raw, "b.pdf"),
|
|
)
|
|
|
|
asyncio.run(main())
|
|
|
|
assert not overlapped, "the inference lock serialised the two reads"
|